Crude Oil Price Time Series Forecasting: A Novel Approach Based on Variational Mode Decomposition, Time-Series Imaging, and Deep Learning

Accurately forecasting crude oil prices is challenging due to market noise and non-stationarity. To address these challenges, we propose a forecasting framework that incorporates variational mode decomposition (VMD), time-series imaging, and bidirectional gated recurrent unit network (BGRU). Our approach eliminates additional assumptions and auxiliary data. First, the raw data are preprocessed through normalization, followed by decomposing multiple stationary sub-series through VMD. Subsequently, three time-series imaging techniques, recurrence plot (RP), Gramian angular field (GAF), and Markov transition field (MTF), are employed respectively to transform the sub-series into two-dimensional images. A convolutional neural network (CNN) is then used to extract features from these images. Finally, the extracted features are fed into BGRU for prediction and the Adam optimizer is used to train models. Experimental evaluations are conducted using a dataset from the U.S. Energy Information Administration (EIA), consisting of the weekly spot prices free on board (FOB) of the West Texas Intermediate (WTI) crude oil, spanning from June 19, 1998, to May 12, 2023. Results demonstrate that all three models constructed following our approach outperform benchmark methods. Specifically, our VMD-RP-BGRU model achieves the best forecasting performance with MAE=2.429, MSE=10.94, MAPE=2.94%, and R-squared=0.9418. The model exhibits reductions of 21.64%, 23.15%, and 36.46% in MAE, MSE, and MAPE, respectively, compared to the seasonal autoregressive integrated moving average (SARIMA) model, and reductions of 21.18%, 22.70%, and 36.08% compared to the Holt-Winters exponential smoothing (HWES) model. Our study contributes to the advancement of crude oil price forecasting techniques and supports informed decision-making in the energy sector.